Interpretability meets medical imaging: shared SAE feature dictionaries and concept-grounded report generation
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored: this snapshot contains no publication counts, venue list, or full output record. This axis would measure total research production — papers per year, venues, and authorship roles — none of which Evolve can currently verify.
Impact—
Not yet scored: no citation counts, h-index, or adoption signals exist in this snapshot. Impact would measure how often these papers are cited and reused across the medical-imaging and interpretability communities; the underlying data is not yet available.
Focus0/100
Scores 0, reflecting a portfolio split evenly — 0.5 versus 0.5 — between medical image analysis / multimodal AI and machine learning interpretability in Evolve's field distribution. The two 2026 papers sit in different directions: CORAL targets interpretable report generation on breast ultrasound and chest X-ray, while SharedSAE builds cross-model feature dictionaries. Both share an interpretability thread, but the record currently reads as two parallel research lines rather than one focused agenda.
Collaboration51/100
Scores 51 (mid-range). Evolve records 12 distinct coauthors, including Xinyue Xu, Hongbin Lin, Weiyang Liu, Adrian Weller, Xiaomeng Li, Kentaro Inui, and Benjamin Heinzerling, a network spanning both of the researcher's active fields. Both cataloged papers list Hu in a middle authorship position with no corresponding-author role, consistent with team-based research rather than solo-led projects.
Activity—
Not yet scored: Evolve has no publication timeline or annual output cadence for this researcher beyond the two 2026 preprints. Activity would measure publishing rhythm and recency of contributions, which cannot be assessed from the current sample.
Reproducibility55/100
Scores 55, averaged over 2 reviewed papers, both rated medium reproducibility readiness. Evolve's review of SharedSAE found it well-controlled and convincing at its scale of four 1B-parameter models, retaining 96.6% of dedicated-SAE variance, while the CORAL medical-report paper was flagged for thin ablations — evidence quality is solid on one line and thinner on the other.
Research summary
Lijie Hu works at the intersection of machine learning interpretability and medical image analysis, a split Evolve's catalog records evenly at 0.5 and 0.5 across two 2026 arXiv preprints. The first, SharedSAE, trains one sparse-autoencoder feature dictionary shared across four 1B-parameter language models, retaining 96.6% of the variance captured by per-model SAEs while yielding transferable feature labels; Evolve's review judged it well-controlled and convincing at that scale. The second, CORAL, attaches concept bottlenecks and SAM3-derived masks to a multimodal large language model for interpretable structured report generation on breast ultrasound and chest X-ray; gains are promising, though templated reports inflate BLEU and ablations remain thin. Both papers list Hu in a middle authorship position without a corresponding-author role, and 12 distinct coauthors appear in the recorded network, including Weiyang Liu, Adrian Weller, Xiaomeng Li, and Kentaro Inui. Radar scores sit mid-range for collaboration (51) and reproducibility (55, averaged over two medium-readiness papers), while focus scores 0 because output is evenly divided between the two fields. Output, impact, and activity remain unscored for lack of citation, venue, and publication-count data. This is an early-stage profile built from a small reviewed sample.
Collaboration pattern
Lijie Hu works in team settings: both cataloged papers list a middle authorship position with no corresponding-author role, and Evolve records 12 distinct coauthors. The network includes Xinyue Xu, Hongbin Lin, Weiyang Liu, Adrian Weller, Xiaomeng Li, Kentaro Inui, and Benjamin Heinzerling, spanning both of the researcher's active fields — a mid-range collaboration score of 51 reflects steady co-authorship rather than a large or dominant network.
Research keywords
Highlights
SharedSAE: one shared feature dictionary across four 1B-parameter language models
CORAL: concept-grounded interpretable report generation for breast ultrasound and chest X-ray
Timeline
- 2026CORAL paper posted to arXiv: concept-grounded reasoning with prompt-driven localization for interpretable structured report generation on breast ultrasound and chest X-ray.
- 2026SharedSAE released on arXiv: one sparse-autoencoder feature dictionary shared across four 1B-parameter language models, retaining 96.6% of dedicated-SAE variance.
- 2026Both preprints enter Evolve's reviewed catalog with medium reproducibility-readiness ratings, forming the basis of this early-stage profile.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Radar scores and bibliometrics come from Evolve's reviewed sample of only 2 papers, not a complete publication record.
- Profile is name-resolved from public bibliographic data; 'Lijie Hu' is a common name, and no institution, country, or ORCID was available to confirm identity.
- Output, impact, and activity dimensions are unscored because citation, venue, and publication-count data are absent from the current snapshot.
- Both cataloged papers are 2026 arXiv preprints; no peer-reviewed venue information is included in this profile.